Logic-RL: Unleashing the reasoning capabilities of large language models with rule-based reinforcement learning. The project aims to reproduce R1 Zero on logic puzzles, using the Knights and Knaves (K&K) puzzle dataset as an example scenario.
Purpose: For research and training experiments on the reasoning capabilities of large language models, to improve the model's logical reasoning performance through rule-based reinforcement learning, and to attempt to reproduce the effects of R1 Zero on logic puzzles.
Main capabilities: The project provides a practical path around data preparation and training execution, and provides a demonstration for generating its own data. At the same time, the repository mentions two related updates: 2025/03/20 released ADORA: a scalable paradigm for guiding learning trajectories; 2025/03/19 for stable length control, refer to https://github.com/lblankl/Short-RL。
Usage: For data preparation, the /data directory can be used directly; if you need to generate data yourself, the project provides a demonstration. For training execution, the project provides relevant instructions for reference implementation.
Applicable scenarios: Suitable for researchers and developers interested in rule-based reinforcement learning, logic puzzle reasoning, and large language model reasoning training, for reproducing experiments and exploring training data preparation and training execution processes.
Limitations: The materials do not provide specific performance metrics, supported model scales, hardware requirements, training duration, detailed dependency environment information, or license terms; apart from the aforementioned ADORA and Short-RL, no more detailed version or pricing information is provided. Some facts are not yet available.
